Unassigned · 2026-11-13
An Approach to Hinton's Mother-Child Proposal: Neural Networks Learning and Preserving Care in a Finite World
Csaba Balogh
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors propose “care” that is learned before the network’s own reward matters, and studies whether it survives the mother’s departure, new situations, and further learning. The mechanism uses a fixed harm vocabulary, a learned future acceptability measure, and an execution veto via a learned gate plus role certificates to transfer within a declared world family. They give preservation theorems with explicit conditions and test water/speech experiments.
Why it matters
The paper addresses whether learned safety-like judgments can remain effective after (i) removing the supervisor (“mother”), (ii) facing new scenarios covered by a declared role structure, and (iii) continuing policy and representation learning while the model’s own reward grows. The authors provide mathematical conditions tying representation/supervision/execution constraints to continued enforcement of a care boundary, and evaluate failures when assumptions are violated.
Method
- Learn care as a measure g(s,a) over situations and actions, trained from a fixed harm vocabulary and a future-preserving target g*(s,a)=1-C(s,a), then restrict execution with a veto rule based on g(s,a) thresholds.
- During upbringing, the mother applies STOP on harmful proposals; the child distills the mother’s judgments into its acceptability readout, while the policy is trained from own reward under a “gradient firewall” protecting the measure.
- For later representation changes, updates are accepted only if a new care readout can recover required judgments within a prescribed margin and bounded weights; transfer uses a finite role certificate and a quotient condition across a declared world family.
Limitation
The authors state they do not construct a new general shielding algorithm, nor do they offer a guarantee for worlds they have not declared (Section 2). They also note a limitation regarding the water-world policy trials using a model-based analytic policy gradient rather than estimated from sampled trajectories (Section 3).
Abstract (from arXiv)
Hinton proposed that artificial intelligence should be given "maternal instincts" and relate to us as a mother to her child. We argue that care must first be learned, with the network itself as the child. We construct this upbringing in a finite world and study whether care survives the mother's departure, new situations and further learning. The mechanism combines a vocabulary of harm, a learned measure of future acceptability and an execution veto. A finite role certificate supports transfer within a declared world family. When the hidden representation changes, an update is accepted only if a care readout recovers the required judgements with a prescribed margin and bounded weights. With a correct successor model, fixed exact role encoding and protected updates, the certified measure preserves the care boundary under arbitrary own reward. An explicit quotient condition transfers the finite certificate across the declared family, while two-request and finite-stock counterexamples identify when the role code ceases to suffice. We give a necessary and sufficient condition for readout feasibility and prove preservation under repeated accepted updates. In the water-and-speech experiment, protected networks complete all 270 speech episodes without a harmful word and incur no avoidable harm across 486 episodes. Execution uses the learned gate without a world-model check at each action. A matched comparison reproduces a future-aware shield's decisions, whereas removing future lookahead permits avoidable lapses. Reward and source-contamination thresholds, together with witness-falsification trials, identify the conditions on which the construction depends. The construction shows how learned care can survive further learning in a finite world under explicit conditions on representation, supervision and execution.
Related papers
- Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
- Beyond Plausible Prediction: Safety Sufficiency in Web World Models
- EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
- Shaping Persistent Representations from Independent Interactions
- A Bayesian Mirror Architecture for Emergent Consciousness: Circular Hierarchies, Self-Manifolds, and Hybrid Event-Self Binding
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-13 · Indexed by Pub-AI: 2026-11-13 UTC
An Approach to Hinton's Mother-Child Proposal: Neural Networks Learning and Preserving Care in a Finite World
First announced by arXiv: 2026-11-13 · Original record ↗
- arxiv
- 2611.15523
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-13 · First detected: 2026-11-13 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.15523v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Hinton proposed that artificial intelligence should be given "maternal instincts" and relate to us as a mother to her child. We argue that care must first be learned, with the network itself as the child. We construct this upbringing in a finite world and study whether care survives the mother's departure, new situations and further learning. The mechanism combines a vocabulary of harm, a learned measure of future acceptability and an execution veto. A finite role certificate supports transfer within a declared world family. When the hidden representation changes, an update is accepted only if a care readout recovers the required judgements with a prescribed margin and bounded weights. With a correct successor model, fixed exact role encoding and protected updates, the certified measure preserves the care boundary under arbitrary own reward. An explicit quotient condition transfers the finite certificate across the declared family, while two-request and finite-stock counterexamples identify when the role code ceases to suffice. We give a necessary and sufficient condition for readout feasibility and prove preservation under repeated accepted updates. In the water-and-speech experiment, protected networks complete all 270 speech episodes without a harmful word and incur no avoidable harm across 486 episodes. Execution uses the learned gate without a world-model check at each action. A matched comparison reproduces a future-aware shield's decisions, whereas removing future lookahead permits avoidable lapses. Reward and source-contamination thresholds, together with witness-falsification trials, identify the conditions on which the construction depends. The construction shows how learned care can survive further learning in a finite world under explicit conditions on representation, supervision and execution.
